National Competence Center – Cybernetics and Artificial Intelligence (03)

Sub-Project name: Artificial Intelligence and Machine Learning
Work package: I. Robotics and Cybernetics for Industry/Society 4.0
ID code: TN01000024/03
Duration: 01/2019 – 12/2022
Principal investigator: Ing. Pavel Burget, Ph.D. (CIIRC CTU)
Participants: Czech Technical University in Prague (CTU)
– Czech Institute of Informatics, Robotics and Cybernetics (CIIRC)
Institute of Computer Science of the Czech Academy of Sciences (ICS CAS)
SIEMENS, s.r.o.
– Application Center (APC) for Production Machines Prague
Results: General-purpose algorithms for machine learning, R, 2020
Software for parts smart bin-picking, R, 2020
Prototype solutions for parts smart bin-picking, Gprot, 2020
Prototype solutions of a force-compliant pick & place delta robot, Gprot, 2020
Use-case demonstrator of picking task – algorithm for robot/ part navigation, R, 2021
Conversational agent for human-robot collaboration, Gprot, 2021+
Python code library, R, 2021+
Robot prototype with diagnostic model, Gprot, 2021+
Text-analysis agent generating formal model from task description, R, 2022
MATLAB/ CUDA code library, R, 2022

Abstract:

Sub-project is focusing on the research and application of ML/AI methods in various fields of industrial automation – grasping and handling, force-compliant manipulation, connection of several machines together and optimal planning of their cooperation.

I. General-purpose library of ML/AI methods

The objective of this area is to design, develop and test algorithms and tools based on a suite of techniques, including deep neural networks, in complex systems to improve their functionality and performance. Their applications will fit into autonomous control systems in the context of Industry 4.0. We will also focus on regression and classification methods for large data sets, regularization techniques for such data, fast classifiers for high-dimension data processing, algorithms for efficient machine learning based on kernel methods, meta-learning algorithms, detection and prevention of adversarial patterns. The results will be in the form of software implemented on a range of architectures, including those dedicated for machine learning problems such as GPUs or specialized neuroprocessors.

II. Grasping and handling of objects

This area covers the field of so-called bin picking and its generalization of handling packs of various shapes and sizes that can be hardly described in an analytical way. Such packs may contain different kinds of parts, whose position in the pack is random. By using machine learning, one can make grasping reliable and independent of the type and shape of the pack content.

In real production environments majority of these cases are represented by picking parts from randomly organized piles, transportation and storage bins, picking from production conveyors, etc.

Instead of relying on having exact models of all packs to compute where to grasp them by the robot we plan that the robot will learn from experience and/or in the way to manipulate the packs by observation of human operators followed by evaluating the success rate and learning from the successful trials.

III. Force-compliant pick & place delta robot

Force-compliant robots typically have six or seven axes, which are connected in an open kinematic chain. Robot can perform its operations, while applying a constant force along the executed trajectory as well as cooperate with a human operator without a danger for the operator to be hurt. The industrial robots with open kinematics are best suited for applications, which require universal behaviour and big working space.

The Pick & Place operations are required to be fast in many cases and cover only a limited workspace.

Since the invention of the delta robot, there have been robot manufacturers offering delta robots typically for pick & place operations. The controllers of such robots are usually closed and do not allow any specific adaptations if they are closer to the physical hardware, i.e. lower-control levels. However, there have also been manufacturers offering just the kinematics equipped with electrical motors to which a standalone motion controller can be connected. This allows designing the controller in a user-specific way that fulfills all necessary requirements. With respect to the objective of this area within NCC, such a lower-level control needs to deal with the force feedback to allow having a force-compliant delta robot.

This project is supported by the Technology Agency of the Czech Republic within the programme National Centres of Competence 1: Support programme for applied research, experimental development and innovation.